Generative Glide - Hit Discovery (Beta) Panel

Set up a Generative Glide workflow that uses generative machine learning models with Glide docking to discover purchasable hit compounds.

To display this panel, do one of the following:
  • click the Tasks button and browse to De Novo Design → Generative Glide (Beta)
  • Overview
  • Using
  • Features
  • Additional Resources

Overview of Generative Glide - Hit Discovery (Beta)

The Generative Glide workflow uses generative machine learning models combined with Glide docking to discover hit compounds from ultra-large chemical libraries. The workflow is designed for libraries containing more than 100 million compounds, such as Enamine REAL and WuXi GalaXi, where conventional docking of every compound is computationally prohibitive.

The workflow consists of four stages:

  1. Prior model training — A recurrent neural network (RNN) learns to generate valid SMILES strings from a large chemical dataset. The prior model captures the general distribution of drug-like chemical space. Pretrained models are provided for major commercial libraries, or you can train a new prior from your own SMILES dataset. Larger datasets generally produce more stable generative models. For custom training, we recommend using 1 million compounds for optimal model performance and stability, a minimum of 50,000 compounds is required.

  2. Supervised model training — A machine learning model is trained using DeepAutoQSAR to predict Glide docking scores from molecular structures. The workflow samples molecules from the prior, docks them with Glide, and trains a surrogate scoring model on those results. This fast surrogate enables rapid evaluation of generated molecules without running full Glide docking on every candidate.

  3. Agent training — An ensemble of reinforcement learning agents fine-tunes the prior model to generate molecules that score well against the supervised model, satisfy property constraints, and dock favorably. Each agent independently explores chemical space, and a diversity filter based on Murcko scaffolds prevents the agents from converging on a narrow set of structures.

  4. Post-processing — Generated molecules are filtered through a Bloom filter to identify purchasable compounds from a commercial library, then docked with Glide to obtain final binding scores. Results are ranked by docking score and analyzed for scaffold diversity. The workflow produces an HTML report summarizing key findings with plots and analysis.

For smaller libraries (fewer than 100 million compounds), consider using the Active Learning Glide Panel instead, which directly screens the enumerated library.

Using the Generative Glide - Hit Discovery (Beta) Panel

The only required input is a Glide receptor grid file, which defines the binding site for docking. Pretrained prior models and Bloom filter binary files for major commercial libraries (Enamine REAL, WuXi GalaXi) are available via secure download links provided at the time of purchase. Contact help@schrodinger.com for help locating them.

To set up a Generative Glide job:

  1. In the Prior model input file section, choose whether to use a pretrained prior model or provide a SMILES dataset to train a new one. Pretrained models are recommended for most use cases.

  2. If you want to identify purchasable compounds, select Use Bloom filter, then choose Local to browse for a Bloom filter binary file (.bin) on your machine, or From server to enter the path to one already on the server where the job runs. Use the Bloom filter for your target commercial library, built from the same dataset used to train the prior model.

  3. Choose the source for the Supervised model input file. The default, Prior model (loaded above), samples molecules from the prior and docks them with Glide to build the surrogate scoring DeepAutoQSAR model automatically. Alternatively, select DeepAutoQSAR to provide your own DeepAutoQSAR model, or Training data to supply a CSV file with pre-computed property values.

  4. In the Glide docking section, browse for the receptor grid file and set the number of CPUs for docking.

  5. In the Agent training section, set the number of reinforcement learning agents and the number of molecules to dock for post-processing validation. Click Define Property Space to set molecular property targets and ranges that constrain the generated chemical space.

  6. Optionally, click the Advanced Settings link to configure excluded substructures, out-of-library sampling, Glide MQ, docking constraints, and other advanced options.

The workflow is fully restartable. If a run is interrupted, resubmit the same job and it resumes from where it left off without re-running completed steps.

Output files

All output is saved to the job name directory. The primary outputs are:

  • Generative_Glide_report_jobname.html — HTML report with workflow status, analysis plots, time breakdown, and summary of key findings.

  • generated_mol_jobname.csv — CSV file containing all generated molecules that passed the filters, with SMILES and target property values, ranked by Glide docking score.

  • Glide pose files — Docking pose output files for visualizing binding poses of generated molecules.

To write out the input file and a script for running the job from the command line, click the arrow next to the Settings button and choose Write. For information on command usage and options, see generative_glide.py Command Help.

Generative Glide - Hit Discovery (Beta) Panel Features

Prior model input file section

Choose the source of the prior model, which is the generative RNN that produces SMILES strings. Select Pretrained model to load an existing model, or select Dataset to train a new prior from a SMILES dataset.

Pretrained model option and Browse button

Select this option to use a pretrained prior model. Browse for a model file (.zip or .gz) or directory containing the model checkpoint. Pretrained models for major commercial libraries are available via secure download links provided at the time of purchase. This is the default selection.

Dataset option and Browse button

Select this option to provide a SMILES dataset (.csv) to train a new prior model. The CSV file must contain a SMILES (case-insensitive) column. Alternatively, provide a directory containing multiple CSV files, which are concatenated.

Use subset option and text box

Select this option to train on a random subset of the dataset. Specify the number of SMILES to use. Larger subsets (at least 50,000 compounds) generally produce more stable generative models. The default is 10,000. Only present when Dataset is selected.

Use Bloom filter option

Select this option to categorize generated molecules against a known compound library using a Bloom filter. A Bloom filter is a compact, probabilistic data structure for fast membership testing that identifies which generated molecules are already present in a commercial chemical library. The Bloom filter should be built from the same dataset used to train the prior model. Bloom filter .bin files and pretrained prior models are typically provided via secure download links sent to the email address used at the time of purchase. Contact help@schrodinger.com for help locating them. Choose Local if the Bloom filter file is on your machine, or From server if it is already on the server where the job runs.

Local option and Browse button

Select this option to use a Bloom filter binary file (.bin) on your local file system, then use the Browse button to locate the file. This is the default.

From server option, text box, and Load button

Select this option when the Bloom filter is already on the server where the job runs, for example on a shared file system that is not accessible from your machine. Enter the absolute path to the Bloom filter binary file (.bin) or to the directory that contains it, then click Load.

Supervised model input file option menu

Choose the source for the supervised model used during agent training. The supervised model is a machine learning surrogate that predicts Glide docking scores, enabling rapid evaluation of generated molecules.

  • Prior model (loaded above) — Samples molecules from the prior and docks them with Glide to build the surrogate scoring DeepAutoQSAR model automatically. This is the default.
  • DeepAutoQSAR — Browse for an existing DeepAutoQSAR Panel model file to use as the surrogate scoring model.
  • Training data — Browse for a .csv file with pre-computed property values to train the surrogate model.
Use single model option

Select this option to train a single supervised model. Not present when DeepAutoQSAR is selected. Not available when Use active learning is selected.

Training sample text box

Specify the number of molecules to sample for training the supervised model. Larger values generally produce better models but increase training time and Glide docking cost. The default is 20,000. Not present when DeepAutoQSAR is selected.

Training time text box

Specify the maximum training time for the DeepAutoQSAR training, in hours. The training stops when this limit is reached and uses the best model found so far. The elapsed time can be significantly longer than the limit specified here, if a model takes a long time to train. The default is 12. Not present when DeepAutoQSAR is selected. Not available when Use single model is selected.

Predicted property option menu

Choose the property column from the training data CSV file to use as the prediction target. Only present when Training data is selected.

Glide docking section

Configure the Glide docking settings used to score generated molecules during supervised model training and post-processing validation.

Grid file file path and Browse button

Browse for the Glide receptor grid file (.zip or .grd) that defines the target binding site. This is the only required input for the workflow.

Number of CPUs text box

Specify the number of CPUs to use for Glide docking. More CPUs run the docking in parallel and finish faster. The number of available CPUs should be at least equal to the number of Glide licenses. The default is 20.

Agent training section

Configure the reinforcement learning agent ensemble. Each agent is a copy of the prior model that is fine-tuned to generate molecules satisfying the specified docking score and property targets. Multiple agents run in parallel, enabling the workflow to explore different regions of chemical space.

Number of agents text box

Specify the number of independent reinforcement learning agents to train in parallel. This controls the throughput of the workflow: more agents will produce a larger number of generated molecules and potentially greater chemical diversity. Agents are designed to run fully in parallel, so increasing the number of agents does not increase wall-clock time as long as enough CPUs are available to run all agents simultaneously. The default is 20.

Molecules to dock text box

Specify the number of generated molecules to select for validation docking with Glide. The default is 20,000.

Define Property Space button

Click this button to open the Target Property Space Dialog Box, where you can set property targets and ranges for generated molecules. Default properties include molecular weight, PSA, AlogP, hydrogen bond donors and acceptors, rotatable bonds, consecutive rotatable bonds, and chiral centers. Additional properties can be added from the Add Properties menu in the dialog.

Use active learning option and Iterations text box

Select this option to run the workflow with active learning, which repeats supervised and agent training over several iterations. After each iteration, the generated molecules that were docked with Glide are added to the supervised model training data, and the models are retrained for the next iteration. Active learning always uses a single supervised model, and the supervised model input defines the model for the first iteration. Specify the number of iterations in the Iterations text box. The default is 3, and the allowed range is 2 to 5.

Advanced Settings link

Click this link to display the Generative Glide - Advanced Settings Pane.

Estimated time to completion text

Displays the estimated wall-clock time for the workflow based on the current settings.

Job toolbar

Manage job submission and settings. See Job Toolbar for a description of this toolbar.

Status bar

Use the Reset button to reset the panel to its default settings and clear any data from the panel. If the panel has a Job toolbar, you can also reset the panel from the Settings button menu.

If you can submit a job from the panel, the status bar displays information about the current job settings and status for the panel. The settings include the job name, task name and task settings (if any), number of subjobs (if any) and the host name and job incorporation setting. The job status can include messages about job start, job completion and incorporation.

The status bar also contains the Help button , which opens an option menu with choices to open the help topic for the panel (Documentation), launch Maestro Assistant, or if available, choose from an option menu of Tutorials. If the panel is used by one or more tutorials, hover over the Tutorials option to display a list of tutorials. Choosing a tutorial opens the tutorial topic.

Tutorials

Related Topics